Despite the development of novel communication tools, emails remain the cornerstone of corporate communication. However, the ubiquity and backward compatibility of the email service makes it a target for malicious attackers. Traditional methods of detecting email threats such as spam include keyword-based filtering, while recent research applies machine learning methods. We argue that email threats can be addressed by combining keyword-based filtering, machine learning, and a collaborative approach, in which organizations pool their resources in federations. We design CyberDART, a system for creating and maintaining federations of organizations. We show how these federations can securely exchange information to create a synergy effect in alleviating email threats. We also show how email contents can be exchanged while providing full anonymity with our PATCH algorithm. To evaluate performance, we implement and deploy CyberDART on virtual servers and use emails from the publicly available Enron database. We demonstrate that CyberDART clearly improves spam detection over traditional methods, e.g., with 20 cooperating organizations, over 50% more spam is detected. This shows that collaborative spam detection based on machine learning can be used successfully provided that security issues are addressed. Also, since CyberDART follows a modular design, it can be extended to address other security threats as well.